Presentation is loading. Please wait.

Presentation is loading. Please wait.

Propagation of Error Berlin Chen

Similar presentations


Presentation on theme: "Propagation of Error Berlin Chen"— Presentation transcript:

1 Propagation of Error Berlin Chen
Department of Computer Science & Information Engineering National Taiwan Normal University Reference: 1. W. Navidi. Statistics for Engineering and Scientists. Chapter 3 & Teaching Material

2 Introduction Any measuring procedure contains error
This causes measured values to differ from the true values that are measured Errors in measurement produce error in calculated values (like the mean) Definition: When error in measurement produces error in calculated values, we say that error is propagated from the measurements to the calculated value Having knowledge concerning the sizes of the errors in measurement → Obtaining knowledge concerning the likely size of the error in a calculated quantity Example 1: measurements: mass, volume Calculation: density=mass/volume measurements: mass, volume of an object Example 2: measurements: length, width of a rectangle Calculation: area=length X width

3 Measurement Error A geologist weighs a rock on a scale and gets the following measurements: None of the measurements are the same and none are probably the actual measurement The error in the measured value is the difference between a measured value and the true value

4 Parts of Error We think of the error of the measurement as being composed of two parts: Systematic error (or bias) Random error Bias is error that is the same for every measurement E.g., a imperfectly calibrated scale always gives you a reading that is too low Random error is error that varies from measurement to measurement and averages out to zero in the long run E.g., Parallax (視差) error The difference in the position of dial indicator when observed from different angles random variable constant constant random variable

5 Two Aspects of the Measuring Process (1/4)
We are interested in accuracy Accuracy is determined by bias The bias in the measuring process is the difference between the mean measurement  and the true value: bias =  - true value The smaller the bias, the more accurate the measuring process Unbiased: the mean measurement is equal to the true value The other aspect is precision Precision refers to the degree to which repeated measurements of the same quantity tend to agree with each other If repeated measurements come out nearly the same every time, the precision is high The uncertainty in the measuring process is the standard deviation  The smaller the uncertainty, the more precise the measuring process

6 Two Aspects of the Measuring Process (2/4)
Example: Figures 3.1 and 3.2 both bias and uncertainty are small bias is large; uncertainty is small bias is small; uncertainty is large both bias and uncertainty are large How about bias? (True value is often unknown in real life) uncertainty is small uncertainty is small uncertainty is large uncertainty is large

7 Two Aspects of the Measuring Process (3/4)
Let be independent measurements, all made by the same process on the same quantity The sample standard deviation s can be used to estimate the uncertainty Estimates of uncertainty are often crude, especially when based on small samples If the true value is know, the sample mean, , can be used to estimate the bias: If the true value is unknown, the bias cannot be estimated from repeated measurements Crude:粗糙的,粗陋的

8 Two Aspects of the Measuring Process (4/4)
From now on, we will describe measurements in the form Where is the uncertainty in the process that produced the measured value Assume that bias has reduced to a negligible level

9 Linear Combinations of Measurements
Each measurement can be viewed as a random variable Its standard deviation represents the uncertainty for it How to compute uncertainties in scaled measurements or combinations of independent measurements? If is a measurement and is a constant, then If are independent measurements and are constants, then

10 Example 3.6 Question: A surveyor is measuring the perimeter of a rectangular lot. He measures two adjacent sides to be  0.05m and  0.08m. These measurements are independent. Estimate the perimeter of the lot and find the uncertainty in the estimate. Answer: Let X = and Y = be the two measurements. The perimeter is estimated by P = 2X + 2Y = m, and the uncertainty in P is P = 2X+2y = So the perimeter is  0.19m.

11 Repeated Measurements
One of the best ways to reduce uncertainty is to take independent measurements and average them (why?) The measurements can be viewed as a simple random sample from a population Their average is therefore the sample mean If are independent measurements, each with mean and standard deviation , then the sample mean, , is a measurement with mean (cf. Sections 2.5 & 2.6) and with uncertainty The uncertainty is reduced by a factor equal to the square root of the sample size

12 Example 3.8 Question: The length of a component is to be measured by a process whose uncertainty is 0.05 cm. If 25 independent measurements are made and the average of these is used to estimate the length, what will the uncertainty be? How much more precise is the average of 25 measurements than a single measurement? Answer: The uncertainty is cm. The average of 25 independent measurements is five times more precise than a single measurement.

13 Repeated Measurements with Differing Uncertainties (1/2)
It happens that the repeated measurements are made with different instruments independently A more suitable way is to calculate a weighted average of the measurements instead of the sample mean of them where

14 Repeated Measurements with Differing Uncertainties (2/2)
Special Case: If and are independent measurements of the same quantity, with uncertainties and , respectively, then the weighted average of and with the smallest uncertainty is given by , where

15 Linear Combinations of Dependent Measurements
If are measurements and are constants, then is a conservative estimate of the uncertainty in For a detailed proof, refer to the textbook (p.170)

16 Example 3.13 Question: A surveyor is measuring the perimeter of a rectanglar lot. Two adjacent sides are measured as  0.05 m and  0.08 m, respectively. These measurements are not necessarily independent. Find a conservative estimate of the uncertainty in the measured value of the perimeter. Answer: Two measurements are denoted by and with uncertainties and , respectively. Let the perimeter be given by . The corresponding uncertainty of is therefore constrained by

17 Uncertainties for (Nonlinear) Functions of One Measurement (1/4)
If is a measurement whose uncertainty is small, and if is a (nonlinear) function of , then This is the propagation of error formula The derivative is evaluated at the observed measurement A constant value! Equation (3.10)

18 Uncertainties for (Nonlinear) Functions of One Measurement (2/4)
Taylor series approximation (linearizing) of constant simply denoted as

19 Uncertainties for (Nonlinear) Functions of One Measurement (3/4)
Definition: If is a measurement whose true value is , and whose uncertainty is , the relative uncertainty in is the quantity However, in practice, is unknown, so the relative uncertainty is estimated by The relative uncertainty is also called the “coefficient of variation”

20 Uncertainties for (Nonlinear) Functions of One Measurement (4/4)
There are two methods for approximating the relative uncertainty in a function 1. Compute using Equation (3.10) and then divide by 2. Compute and use equation Equation (3.10) to find , which is equal to Note that relative uncertainty is a number without units It is frequently expressed as a percent

21 Example 3.14 Question: The radius of a circle is measured to be cm. Estimate the area of the circle and find the uncertainty in this estimate Answer: The area is given by The estimate of the area is cm The uncertainty of the area is So the estimate of the area of the circle can be expressed by

22 Example 3.17 Question: The acceleration for a mass down a frictionless inclined plane is given is the acceleration due to gravity (the uncertainty in is negligible) is the angle of inclination of the plane ( ) Find the relative uncertainty in Answer: Rad=radian弧度

23 Uncertainties for Functions of Several Independent Measurements
If are independent measurements whose uncertainties are small, and if is a function of , then This is the multivariate propagation of error formula In practice, we evaluate the partial derivatives at the point

24 Uncertainties for Functions of Several Dependent Measurements
If are not independent measurements whose uncertainties are small, and if is a function of , then a conservative estimate of is given by Because in most cases of practical applications, the covariance between dependent measurements are unknown This inequality is valid in almost all practical situations; in principle it can fail if some of the second partial derivatives of U are quite large Due to the linear approximation using Taylor series

25 Example Question: Two perpendicular sides of a rectangle are measured to be cm and cm. Find the absolute uncertainty in the area ( and are known independent) Answer: First, we need the partial derivatives: and , so the absolute uncertainty is

26 Summary We discussed measurement error
Then we talked about different contributions to measurement error We looked at linear combinations of measurements (independent and dependent) We considered repeated measurements with differing uncertainties The last topic was uncertainties for functions of one measurement


Download ppt "Propagation of Error Berlin Chen"

Similar presentations


Ads by Google